Kimi K2.5 and moonshot-v1 Sunset Today: What Migrating to kimi-k3 Actually Costs

Two Moonshot AI API generations reach end-of-life today, August 31: kimi-k2.5, the multimodal flagship that defined the K2 era, and the entire moonshot-v1 series — the context-tiered models (8k/32k/128k/auto) whose per-window pricing was Moonshot’s original signature. Integrations still calling either will start failing as the platform completes its sunset; the official migration target is kimi-k3.

What Is Being Retired

According to Moonshot’s official model documentation, both retirements were announced in July and proceeded in two phases: kimi-k2.5 and moonshot-v1 stopped accepting new users after the Kimi K3 launch, and today the full platform sunset takes effect.

kimi-k2.5 — Moonshot’s first trillion-parameter multimodal model, with native text/image/video input, thinking and non-thinking modes, and multi-sub-agent orchestration. Also going dark on third-party surfaces: Tencent Cloud announced it would take K2.5 offline at 00:00 Beijing time on August 31.
moonshot-v1 (8k/32k/128k/auto) — the legacy family that priced the same model differently by context window, from $0.20/$2.00 per million tokens at 8k up to $2.00/$5.00 at 128k. It is the mechanic being retired as much as the models themselves.

The Migration Target: kimi-k3

K3, open-sourced in July, is a generational jump rather than a drop-in replacement: 2.8 trillion parameters built on the KDA hybrid linear attention mechanism, native visual understanding, and a context window expanded to 1,048,576 tokens — the first open-source model at that parameter scale, per Moonshot’s documentation. The K3 rate card is also the company’s first genuinely premium one: $3.00 per million input tokens and $15.00 per million output.

The Migration Is Not Price-Neutral

This is the part that will surprise teams budgeting mechanically. Per current pricing analysis, where you land depends entirely on which v1 tier you’re leaving:

moonshot-v1-128k users ($2.00/$5.00) migrating to Kimi K2.5’s replacement tier ($0.60/$3.00) actually get a substantial cut — with double the context window.
moonshot-v1-8k users ($0.20/$2.00) lose the cheapest input rate on the platform and land on a rate roughly 3× higher.
Long-context RAG workloads moving to K3 get the real cost shock: a representative workload (60M input tokens monthly at a 60% cache-hit rate, 15M output) runs roughly $63/month on K2.5-class pricing but approximately $308/month on K3 — with output alone accounting for about $225 of it.

The K2.5-to-K3 jump is also architectural, not just commercial: K2.5 used traditional attention; K3’s KDA hybrid linear attention is what makes the 1M-token window feasible. Code that worked against kimi-k2.5 should be re-tested against K3’s thinking/non-thinking modes and updated visual-input formats rather than assumed compatible.

What Developers Should Do Now

Audit your model strings. Anything referencing kimi-k2.5 or moonshot-v1-* breaks today. Inventory routes, prompts, and eval harnesses that hardcode them.
Re-benchmark before re-pointing. If you were on v1-8k for cost, compare K2.5-tier and K3 pricing on your actual traffic before choosing; the answer differs by workload.
Exploit the context window. Workloads that paid for v1-128k tiering to fit long documents can consolidate on K3’s flat 1M context — fewer chunking workarounds, simpler retrieval.
Check third-party hosts separately. Tencent Cloud and other providers run their own timelines (Azure reportedly holds K2.5 access into 2027); a migration off Moonshot’s platform doesn’t automatically mean off every surface.

What Remains Unclear

Moonshot’s documentation confirms the sunset date and the K3 migration target, but details like exact error behavior for in-flight requests after today, and whether any grace period applies to enterprise contracts, are not publicly documented. Teams with production traffic on either model should verify behavior directly rather than assume a hard cutoff at a specific hour.

Conclusion

Today closes out the pricing model that made Moonshot’s name — context-as-a-meter — along with the K2.5 flagship. The K3 destination is technically superior on every axis Moonshot advertises, but the bill changes unevenly: some v1 users cut costs by migrating, others face 3–5× effective increases depending on tier and workload. The sunset deadline is real, and the pricing math is per-workload. Do it before your integration discovers it for you.